JevLoop (parkavenue9639)
Python agent runtime where Jev Choice decisions pick tools and arguments, and low-confidence picks go to an LLM for review.
Overview
JevLoop (parkavenue9639) is an open-source agent runtime in Python. Its loop is observe, Jev decide, guard, execute, record in a ledger and repeat. Jev handles bounded tool decisions as Choice questions, read as purpose, then tool, then arguments; an LLM is called only when language must be written or when Jev's confidence is low and a choice needs review. Deterministic code owns execution, safety, durability and state, and the runtime supports isolated Docker workspaces.
Candidate questions are rebuilt each iteration from tool results, so a file written in one step becomes an argument option for READ_FILE in the next; the README shows the pool growing from 8 to 14 questions within one task. A console visualises candidates, consumed choices and LLM hand-offs. The runtime also supports paired runs against an LLM-only agent; in one selected historical run JevLoop finished in 26.4 s and 5 steps versus 64.1 s and 14 steps, which the README says is not a general performance claim. It is not an official TypeSafe product.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What is the purpose of the next step?(inferred) | — |
| Choice | Which tool should be called next? | — |
| Choice | Which argument (e.g. which file) should the tool use? | — |
How it works
- Observe the transcript and tool results to build candidate questions
- Jev chooses purpose, tool and arguments from the candidate pool
- Low-confidence choices go to an LLM for review; text content is authored by the LLM
- Guard checks the call, then it executes in the workspace and is recorded in the ledger
- Repeat until the task completes
Strengths
- Bounded choices skip full LLM calls on the fast path
- Deterministic guard and ledger own execution and state
- Docker isolation and paired LLM-only comparisons built in
- Bilingual docs (English and Chinese)
Caveats
- Published comparison is a single selected run
- Still relies on an LLM for authored content and reviews
- Independent project, not official TypeSafe software
At a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- agent runtime with web console, Docker sandboxes
- Has code
- yes
Use cases and integrations
How curated lists describe it
Agent runtimes: a Python runtime where Jev `Choice` decisions select tools and targets, uncertain decisions escalate to an LLM, and a shared guarded kernel supports isolated Docker workspaces and paired LLM-only comparisons.
yibie/awesome-jevPython agent runtime where Jev `Choice` decisions select tools and targets, uncertain decisions escalate to an LLM, and a shared guarded kernel supports isolated Docker workspaces and paired LLM-only comparisons.
cobanov/awesome-jevA Python agent runtime powered by TypeSafe Jev: guarded tool execution, isolated Docker sandboxes, and side-by-side LLM comparisons.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide in JevLoop?
The purpose, tool and arguments of each next step, as Choice questions over a candidate pool rebuilt every iteration.
When is the LLM used?
When content must be written, or when Jev's confidence is low and its choice needs review.
Is it faster than an LLM-only agent?
One published paired run was faster and shorter, but the README says it is not a general claim.
Related projects
One skill and MCP server for Claude Code, Codex and Hermes where Jev gates large-model turns, routes models and guards Git steps.
MCP serverbetaPythonPython harness for open-ended goals: AI-authored tools and checks, Jev choosing the next tool and arguments from fresh state.
ProjectbetaPythonDecision middleware for AI agents that uses local Laya to pick tools and judge results, falling back to an LLM when unsure.
ProjectbetaPythonSelf-improving multi-platform agent harness with an optional Jev companion for typed judgments and per-session model routing.
ServicebetaTypeScriptOmO/senpi agent plugin where Jev or Span-01 rates which skills and next tools fit the task and spots loops or completion.
Agent skillbetaTypeScriptClaude Code function hook that sends the agent's this-or-that questions to Jev and optionally ranks installed skills per prompt.
ProjectbetaTypeScript